How to optimize neural networks for over-dispersed target data?

I am training a feedforwardnet that ought to predict accident rates from traffic and road characteristics of a large number of road segments. The target data is highly over-dispersed (a lot of zeros, where no accidents occurred over the observed period). I guess that this might be one problem. Does anyone have experience in optimizing network parameters such as training algorithm, number of neurons, even net type? A couple of pointers would be much appreciated.

Answer 1

[Greg Heath]

Use a radial basis function net

 help newrb
 doc newrb

Also consider newrbe , newgrnn, newpnn, etc